My research focuses on reducing intelligent models’ dependence on human-generated data. Weakly supervised learning provided the initial point of departure: how can effective learning signals be derived from limited annotations? As this research progressed, the question extended from the data space to the weight space. Through model merging, multi-task fusion, weight generation, and model editing, my work explores how to reuse, combine, and update the knowledge and capabilities already encoded in models. This trajectory has naturally expanded toward large language models and foundation models.https://tianbowen.net/zh-CNhttps://tianbowen.net/blog/test/https://tianbowen.net/blog/test/大海捞针（needle-in-a-haystack）这类评测已经饱和了，但它测的其实是检索，不是推理。Mon, 27 Jul 2026 00:00:00 GMT评测长上下文https://tianbowen.net/blog/attention-sinks/https://tianbowen.net/blog/attention-sinks/从 KV cache 中丢掉第一个 token，模型质量的下降远超它的语义信息量所能解释的程度。Mon, 20 Jul 2026 00:00:00 GMT可解释性注意力机制